J. Mol. Biol. (2008) 377, 914–934
doi:10.1016/j.jmb.2008.01.049
Available online at www.sciencedirect.com
Rescoring Docking Hit Lists for Model Cavity Sites:
Predictions and Experimental Testing
Alan P. Graves 1,2 †, Devleena M. Shivakumar 3 †, Sarah E. Boyce 1,4 ,
Matthew P. Jacobson 1 ⁎, David A. Case 3 ⁎ and Brian K. Shoichet 1 ⁎
1
Department of Pharmaceutical
Chemistry, University of
California, San Francisco,
1700 4th Street, San Francisco,
CA 94158-2330, USA
2
Graduate Group in Biophysics,
University of California,
San Francisco, 1700 4th Street,
San Francisco, CA 94158-2330,
USA
3
Department of Molecular
Biology, The Scripps Research
Institute, 10550 N. Torrey Pines
Road, La Jolla, CA 92037, USA
4
Graduate Group in Chemistry
and Chemical Biology,
University of California,
San Francisco, 1700 4th Street,
San Francisco, CA 94158-2330,
USA
Received 21 September 2007;
received in revised form
12 January 2008;
accepted 17 January 2008
Available online
30 January 2008
Molecular docking computationally screens thousands to millions of
organic molecules against protein structures, looking for those with
complementary fits. Many approximations are made, often resulting in
low “hit rates.” A strategy to overcome these approximations is to rescore
top-ranked docked molecules using a better but slower method. One such is
afforded by molecular mechanics–generalized Born surface area (MM–
GBSA) techniques. These more physically realistic methods have improved
models for solvation and electrostatic interactions and conformational
change compared to most docking programs. To investigate MM–GBSA
rescoring, we re-ranked docking hit lists in three small buried sites: a
hydrophobic cavity that binds apolar ligands, a slightly polar cavity that
binds aryl and hydrogen-bonding ligands, and an anionic cavity that binds
cationic ligands. These sites are simple; consequently, incorrect predictions
can be attributed to particular errors in the method, and many likely ligands
may actually be tested. In retrospective calculations, MM–GBSA techniques
with binding-site minimization better distinguished the known ligands for
each cavity from the known decoys compared to the docking calculation
alone. This encouraged us to test rescoring prospectively on molecules
that ranked poorly by docking but that ranked well when rescored by MM–
GBSA. A total of 33 molecules highly ranked by MM–GBSA for the three
cavities were tested experimentally. Of these, 23 were observed to bind—
these are docking false negatives rescued by rescoring. The 10 remaining
molecules are true negatives by docking and false positives by MM–GBSA.
X-ray crystal structures were determined for 21 of these 23 molecules. In
many cases, the geometry prediction by MM–GBSA improved the initial
docking pose and more closely resembled the crystallographic result; yet in
several cases, the rescored geometry failed to capture large conformational
changes in the protein. Intriguingly, rescoring not only rescued docking
false positives, but also introduced several new false positives into the topranking molecules. We consider the origins of the successes and failures in
MM–GBSA rescoring in these model cavity sites and the prospects for
rescoring in biologically relevant targets.
© 2008 Published by Elsevier Ltd.
Edited by B. Honig
Keywords: decoys; molecular docking; virtual screening; MM–GBSA; cavity
*Corresponding authors. E-mail addresses: matt.jacobson@ucsf.edu; case@scripps.edu; shoichet@cgl.ucsf.edu.
† A.P.G. and D.M.S. contributed equally to this work.
Present address: D. M. Shivakumar, Institute for Molecular Pediatric Sciences, University of Chicago, 929 E. 57th Street,
Chicago, IL 60637, USA.
Abbreviations used: L99A, Leu99 → Ala mutant of T4 lysozyme; L99A/M102Q, Leu99 → Ala and Met102 → Gln double
mutant of T4 lysozyme; CCP, Trp191 → Gly mutant of cytochrome c peroxidase; MM–GBSA, molecular mechanics with
generalized Born surface area approximation; PLOP, Protein Local Optimization Program; ACD, Available Chemicals
Directory; PDB, Protein Data Bank; SGB, surface generalized Born.
0022-2836/$ - see front matter © 2008 Published by Elsevier Ltd.
Rescoring Docking Hit Lists
Introduction
Molecular docking computationally screens large
databases of small molecules against a macromolecular binding site of defined structure. The technique is often used to find novel ligands for drug
discovery. Notwithstanding important successes,1–8
docking continues to struggle with many methodological deficits. Many approximations are made to
screen many molecules in a timely fashion. These
include using only one conformation of the protein,
neglecting the internal energies of the docking
molecules, using simplified models of ligand solvation energies, typically ignoring protein desolvation,
and ignoring most entropic terms entirely. These
and other shortcuts lead to the high false-positive
and false-negative rates for which docking screens
are notorious. Docking methods are unreliable for
affinity prediction and, except in domains of highly
related compounds, even for rank ordering the
likely hits that emerge from the virtual screens.
To overcome these deficits, several groups have
combined disparate scoring functions in a consensus
fashion to capitalize on the strengths and overcome
the deficiencies of individual methods.9–12 This
“consensus scoring” approach is attractive when it
has worked, but its theoretical underpinnings are
slim.13 An alternative approach involves using a
higher level of theory to rescore the docking hit lists
after the docking calculation has completed. The goal
is to reevaluate the top docking hits for energetic
complementarity to the target after including more
terms and degrees of freedom than modeled by the
docking program. Because more terms are considered, rescoring is typically much slower than docking, so much so that only the top-scoring docking
pose of the best scoring docked molecules are often
considered. This approach has been adopted by
versions of the program GLIDE.14 Here ligands are
first docked using simplified and relaxed criteria
and are then refined by more sophisticated and
stringent evaluation of the energies of binding.
Similarly, Wang et al. used a hierarchical technique
that begins with initial database screening and progresses to molecular mechanics–Poisson–Boltzmann
surface area (MM–PBSA) rescoring to find HIV-1
reverse transcriptase inhibitors. 15 The combination of an initial docking screen with subsequent rescoring by a molecular mechanics–generalized Born
surface area (MM–GBSA) method has been used to
improve enrichment of known ligands for several
enzymes in retrospective studies and even to identify substrates.16–20
Such MM–PBSA and MM–GBSA methods involve
minimization and often dynamic sampling of the
protein–ligand complexes, and include ligand and
receptor conformational energies and strain. They
evaluate the electrostatics and solvation components of the binding energy by PB or GB methods,
including both ligand and receptor desolvation.
The MM–GBSA binding energy is determined by
Ecomplex − Ereceptor − Eligand where E is an MM–GBSA
estimate and solute configurational entropy effects
915
are ignored. In this article, we focus on relative
binding energies of different ligands to the same
receptor, so the free receptor energy (Ereceptor) does
not affect the results. Because the MM–GBSA
function includes both internal energies and solvation free energies, and because we explicitly subtract
complex (Ecomplex) and ligand (Eligand) contributions,
desolvation effects upon complex formation for both
the ligand and the receptor are included, at least in
principle. There are three main limitations: (1) the
force fields and solvation energies are not uniformly
accurate; (2) for reasons of computational efficiency,
only a small part of configuration space near the
DOCK starting pose is really explored; and (3) configurational entropy effects are ignored. Notwithstanding these limitations, the MM–GBSA methods
represent a substantially higher level of theory than
that encoded by most docking programs and are
attractive alternatives to a more complete treatment
of the energies of interaction by free-energy perturbation and thermodynamic integration,21 which
remain the gold standard but are very slow.
In this study, we set out to test MM–GBSA
rescoring of docking hit lists in simple model cavity
sites. These sites have been engineered into the
buried cores of proteins and bind multiple small organic molecules. In contrast to most drug targets,
these cavities are small (150–180 Å3), buried from
bulk solvent, and are dominated by a single interaction term. The L99A (Leu99 → Ala) cavity in T4 lysozyme22 is almost entirely apolar, the L99A/M102Q
(Leu99 → Ala/Met102 → Gln)23 cavity in the same
protein has a single hydrogen-bond acceptor (the introduced Gln102), whereas the W191G (Trp191 →
Gly) cavity in cytochrome c peroxidase (CCP)24,25
has a single anionic residue, Asp235 (Fig. 1). The
ligands recognized by these sites correspond to these
features: the hydrophobic L99A binds small, typically aromatic nonpolar molecules; the slightly polar
L99A/M102Q binds not only both apolar molecules
but also those bearing one or two hydrogen-bond
donors, whereas the anionic W191G cavity almost
exclusively binds small monocations. The simplicity
of these sites is conducive to disentangling the energetic terms of ligand binding, which are so often
convoluted in drug targets with their larger, more
complex binding sites. It should be noted that previous work with solvent-exposed sites has suggested
that a major advantage of MM–GBSA scoring functions is calculating partial receptor desolvation upon
ligand binding.17 This benefit with complex solventexposed binding sites may be less relevant in the
buried cavity sites, especially the hydrophobic L99A
and polar L99A/M102Q sites, which are mostly
desolvated. (It is our experience that the cavity sites,
in fact, impose a greater strain on the GBSA solvent
models to fully desolvate the pockets.)
In the cavity sites, as in other simplified sites,27 an
incorrect prediction is often informative, identifying
a single problematic term in a scoring function;
we have used these cavities as model binding sites
to identify problems in molecular docking23,28–30
and, more recently, thermodynamic integration.21
Rescoring Docking Hit Lists
916
Fig. 1. The model cavity sites. (a) Cavity binding site in
T4 lysozyme L99A with benzene bound. (b) Cavity
binding site in T4 lysozyme L99A/M102Q with phenol
bound; the hydrogen bond with the Oε2 oxygen of Gln102
is represented by a dashed line. (c) Cavity binding site of
cytochrome c peroxidase W191G with aniline bound; the
hydrogen bond with Asp235 is represented by a dashed
line. The heme and an ordered water molecule are also
depicted. In (a), (b), and (c), the cavities are represented by
a tan molecular surface and the protein ribbons are green.
Rendered with the program PyMOL.26
Others have found them attractive test systems for
method development studies.31–34 An important
advantage of these cavity sites is that they are experimentally tractable for detailed, prospective testing
of ligand predictions. Because the ligands they bind
are small—in the 70- to 150-amu range—many
possible ligands are readily available commercially,
which is rarely true of drug targets.35 The binding of
these predicted ligands may be tested by direct
binding assays, and the structures of the ligand–
protein complexes may be routinely determined by
X-ray crystallography to resolutions better than 2 Å.
Extensive study in the Matthews, Goodin, and our
own laboratories has resulted in many tens of
diverse ligands for each cavity, as well as tens of
“decoys,” which are molecules that were predicted
to bind to the sites but for which no binding was
observed at concentrations as high as 10 mM on experimental testing.21,23,28–30
We thus used these three simple model cavity sites,
L99A, L99A/M102Q, and W191G, as templates to
measure the strengths and weaknesses of MM–
GBSA rescoring of docking hit lists. We used two
rescoring programs: Protein Local Optimization
Program (PLOP),36,37 with binding-site side-chain
rotamer search and minimization, and AMBERDOCK, using short molecular dynamics (MD) steps
and minimization of binding-site residues (Materials
and Methods). Molecular docking was used to
screen compound libraries that contained between
5000 and 60,231 fragment-like molecules from the
Available Chemicals Directory (ACD); the library
size was chosen to partly mitigate issues of size and
charge bias from the library alone and to be
consistent with earlier studies in these sites (Results).28,29 The single best pose for each compound
that ranked among the top 5000 or 10,000 compounds by docking was then rescored by both MM–
GBSA programs. Multiple known ligands and
decoys were among the molecules rescored for all
three sites' rescored sets. In retrospective calculations, MM–GBSA rescoring improved the separation of ligands from decoys in each of the cavities.
We then tested 33 new ligands that were predicted
to bind by the MM–GBSA methods that docking
alone ranked poorly, generally much worse than
the top 500. To investigate the detailed basis of
the MM–GBSA predictions, we determined crystal
structures for 21 of these new ligands and compared them to the geometries predicted by theory.
These studies suggest areas where MM–GBSA methods can contribute to the success of virtual screening and areas where this method faces important
challenges.
Results
Retrospective docking and rescoring in the
hydrophobic cavity
Approximately 60,000 small molecules were
docked into the hydrophobic cavity L99A using
DOCK3.5.5423,38 (Fig. 1a). The compounds in this set
were selected from a much larger library so as not to
exceed 25 non-hydrogen atoms, as previously described.29 This reduced the enrichment-factor bias
that would have otherwise occurred by the trivial
ability of the docking program to remove com-
Rescoring Docking Hit Lists
pounds that were simply too large to fit in the cavities. We note that reducing the number of molecules
to 60,000 from the several million that are in the ACD
or ZINC39 databases has the effect of reducing our
enrichment factors. Among the top-scoring 10,000
molecules were 39 known ligands and 40 experimentally tested decoys. DOCK found 44% (17 molecules) of these ligands and 43% (17 molecules) of
these decoys among the top 500 molecules (Fig. 2a).
Ligands such as toluene (DOCK rank 32), benzene
(DOCK rank 151), and ethylbenzene (DOCK rank
301) are small, aromatic, and hydrophobic compared
to known decoys such as nitrosobenzene (DOCK
rank 125), phenol (DOCK rank 234), and 3-methylpyrrole (DOCK rank 435). Like the ligands, these
decoys are also small and aromatic, but are presumably too polar for the hydrophobic cavity to overcome their desolvation penalty (Fig. 1a).
The top-ranking 10,000 docking hits for the
hydrophobic cavity were re-ranked by PLOP and
the top-ranking 5000 docking hits were re-ranked
by AMBERDOCK; fewer molecules were treated
by AMBERDOCK simply because it was much more
computationally intensive than PLOP. For both methods, the enrichment of the ligands actually decreased slightly relative to that achieved by docking
alone; that is to say, fewer ligands were found among
the very best scoring molecules (Fig. 2a). Rescored by
PLOP, 41% (16 molecules) of the known ligands were
found among the top 500 molecules, whereas 28%
(11 molecules) were found by AMBERDOCK. Both
enrichment factors were lower than those found by
docking alone. On the other hand, the enrichment
of the known decoys was lower still (Fig. 2a). Only
5% of the decoys (2 molecules) were ranked among
the top 500 molecules by PLOP and only 13% (5
molecules) were so ranked by AMBERDOCK. This
represents a substantial improvement on docking
alone, one that reflects a significant change in the
relative energies of the ligands and decoys. For instance, in the L99A cavity the average differential
energy between the first 10 ligands and the first 10
decoys was only 0.7 kcal/mol by docking. Meanwhile, the average total energy for the top 10 docked
ligands was −15.8 kcal/mol and the difference between the 1st and the 10th ranked ligand is 2.9 kcal/
mol; the ligands and decoys were essentially indistinguishable by docking energy. For the PLOP
rescored molecules, conversely, the average difference in energies for the top 10 ligands and decoys was
4.0 kcal/mol. Meanwhile, average energy for the top
10 ligands was −21.7 kcal/mol and the difference
between top ranked ligand and the 10th was 5.5
kcal/mol; the best ligands and decoys are separated
significantly by rescored energy. We should note that
both the ligand enrichment and the decoy enrichment
are strongly biased for docking—many of the ligands
and almost all of the decoys were originally tested
based on docking predictions23,29,30—so it is reasonable to expect that the enrichment of ligands will be
higher by docking, as will the decoys. Perhaps more
informative then is the separation of the ligands
from the decoys, as measured by the ratios of their
917
Fig. 2. Retrospective enrichment of ligands and decoys
for (a) the hydrophobic L99A cavity, (b) the polar L99A/
M102Q cavity, and (c) the anionic W191G cavity. The plots
depict the percentage of known ligands (continuous lines)
or decoys (dashed lines) found (y-axis) at each percentage
level of the ranked database using the top 10,000 best
scoring docking hits (x-axis) for L99A (a) and L99A/
M102Q (b) and the 5400 best scoring docking hits (x-axis)
for CCP (c). Docking enrichment of known ligands
(continuous lines) and decoys (dashed lines) are represented by the dark blue curves. PLOP enrichment of
known ligands (continuous lines) and decoys (dashed
lines) are represented by the pink curves. AMBERDOCK
enrichment of known ligands (continuous lines) and
decoys (dashed lines) are represented by green curves.
enrichment factors. These were improved eightfold
by PLOP and twofold for AMBERDOCK, relative to
that of DOCK in this hydrophobic cavity.
918
Retrospective docking and rescoring in the
polar cavity
The same 60,000 molecules were docked into the
polar cavity L99A/M102Q (Fig. 1b). Among the topscoring 10,000 molecules were 58 ligands and 17
experimentally tested decoys. DOCK found 45% (26
molecules) of these ligands and 35% (6 molecules) of
these decoys among the top 500 molecules (Fig. 2b).
The increased polarity from Oε of the Gln102 side
chain in the cavity accommodates the binding of
phenol (DOCK rank 354) and 3-methylpyrrole (DOCK
rank 307), which are decoys for the L99A cavity, as
well as hydrophobic ligands such as toluene (DOCK
rank 16) and benzene (DOCK rank 78). The increased
polarity of the site only goes so far, however, and it
cannot accommodate decoys such as 1-vinylimidazole (DOCK rank 136) or 2-aminophenol (DOCK
rank 208), whose polarity is presumably still too great
for the single carbonyl oxygen of the site to overcome
the attendant desolvation terms.
The top 10,000 docking hits for the polar cavity
were re-ranked by PLOP and the top 5000 re-ranked
by AMBERDOCK. For both methods, the enrichment of the ligands again decreased slightly relative
to the docking enrichment factor (Fig. 2b). Rescored
by PLOP, 22% (13 molecules) of the known ligands
were found among the top 500 molecules, whereas
34% (20 molecules) were found by AMBERDOCK.
However, the enrichment of the known decoys was
lower still. None of the decoys were ranked among
the top 500 molecules by PLOP or AMBERDOCK,
in contrast to DOCK where 35% (6 molecules) of
the known decoys were scored among the top 500
molecules. As in the hydrophobic site, despite the
decrease in overall ligand enrichment, the separation of the ligands from the decoys was improved
substantially for the polar cavity: by 20-fold for
PLOP and 35-fold for AMBERDOCK.
Retrospective docking and rescoring in the
anionic cavity
Approximately 5400 molecules were docked in the
charged cavity of CCP (Fig. 1c). This library was also
selected from a much larger set to reduce enrichmentfactor bias from trivial physical noncomplementarity
between library molecules and the CCP cavity.28
Thus, any molecules from the larger ACD that had
unfavorable van der Waals scores (i.e., simply did not
fit), or that bore an anionic charge, were removed
from the larger library. As with the lysozyme cavities,
this smaller library of more physically plausible ligands reduces the enrichment factors we would
otherwise achieve with docking. Within this database
were 40 known ligands and 20 experimentally tested
decoys. DOCK found 78% (31 molecules) of these
ligands and 20% (4 molecules) of these decoys among
the top 500 molecules (Fig. 2b). The anionic cavity
typically binds cationic ligands such as 2-aminopyridine (DOCK rank 6) and imidazole (DOCK rank
227). Most neutral polar compounds, such as 3,5difluoroaniline (DOCK rank 148), and apolar com-
Rescoring Docking Hit Lists
pounds, such as toluene (DOCK rank 411), are decoys for this cavity, as are anionic compounds or
those bearing a formal charge greater than +1.
All of the 5400 docking hits for the anionic cavity were re-ranked by PLOP and AMBERDOCK.
Rescored by PLOP, 83% (33 molecules) of the known
ligands were found among the top 500 molecules,
and 80% (32 molecules) were found by AMBERDOCK (Fig. 2c). Both enrichment factors are comparable to those found by docking alone, which
found 83% (33 molecules) of the known ligands
among the top 500 molecules. On the other hand,
fewer of the known decoys were enriched by the
MM–GBSA methods. None of the known decoys
were ranked among the top 500 molecules by PLOP
or AMBERDOCK, and the best scoring decoy
ranked 655 for PLOP and 785 for AMBERDOCK
compared to 145 for docking. Thus, whereas the
overall enrichment of the ligands relative to the rest
of the database molecules remained unchanged, the
separation of the ligands from the decoys was
improved by fourfold for PLOP and AMBERDOCK.
Prediction and experimental testing of new
ligands
A more robust test, one less biased by previous
knowledge, involves prospective prediction of new
ligands. For each of the three cavities, we looked for
molecules that had been poorly ranked by docking
but that ranked well by either PLOP or AMBERDOCK or both. We note that our use of “well” and
“poorly” ranked is inexact because there is no fully
reliable way to separate molecules based on docking
energies alone. We therefore looked for molecules
where the rankings changed substantially—typically rising from ranks lower than 1500 to ranks in
the top 200. Of the 33 molecules selected, 24 were
ranked worse than 1500th by docking, 7 were
ranked between 500 and 1500, and 2 were ranked
between 300 and 500. The rankings of all 33 rose to
be among the top 200 on rescoring. Our choice of 200
was purely pragmatic, as it is a reasonable number
of top-ranking hits to visualize and consider for
testing, which is often done when picking docking
hits; another reasonable cutoff would have been top
500. Nine compounds were picked and tested for the
hydrophobic L99A cavity, 10 were tested for the
polar L99A/M102Q cavity, and 14 were tested for
the anionic W191G cavity. Structures for 21 of these
33 molecules in complex with the cavities were determined by protein crystallography, allowing us to
compare the predicted and experimental geometries
in detail. In the following discussion, we report
whether binding was detected at a single concentration tested. The actual affinities were not measured but will often be substantially better than the
concentration reported.
New L99A ligands predicted by rescoring
All of the nine ligands predicted by PLOP and
AMBERDOCK were relatively large compounds
Score and ranka
Structure
b
c
d
ΔΔH
(kcal/mol)
ΔTm
(°C)
Binding
detected
Structure
determined
10
3.0
31.0
6.5
Yes
Yes
− 16.22 (1243)
b10
3.0
6.2
1.3c
Yes
Yes
− 7.69 (5290)
10
6.8
1.6
− 0.9c
No
No
b10
3.0
10.0
1.6c
Yes
Yes
5
3.0
12.0
1.2
Yes
Yes
− 25.19 (27)
10
3.0
8.1
1.5
Yes
Yes
NRd
−25.17 (28)
10
3.0
5.1
1.5
Yes
No
NRd
−24.92 (30)
10
3.0
14.9
2.8
Yes
No
−19.64 (55)
−23.13 (68)
10
3.0
2.7
0.0
No
No
DOCK
AMBER
β-Chlorophenetole (1)
− 4.89 (3786)
−22.38 (5)
− 26.31 (15)
4-(Methylthio)nitrobenzene (2)
− 5.69 (3358)
−22.36 (6)
1-Phenyl-semicarbazide (3)
− 4.49 (3965)
−22.03 (8)
2,6-Difluorobenzyl bromide (4)
− 10.59 (1046)
−22.01 (9)
− 21.10 (186)
2-Ethoxyphenol (5)
− 6.74 (2806)
−21.54 (12)
− 15.19 (1642)
3-Methyl benzylazide (6)
− 10.54 (1061)
−19.58 (57)
cis-3-Hexenyl formate (7)
3.61 (7746)
6-Methyl-1,5-heptadiene (8)
1.78 (7035)
2-Phenoxyethanol (9)
a
pH
Compound (ID)
−5.76 (3323)
PLOP
|C|b (mM)
Rescoring Docking Hit Lists
Table 1. Compounds predicted by AMBERDOCK and PLOP to bind to T4 lysozyme L99A
Compound scores and ranks (in parenthesis) for DOCK, AMBERDOCK, and PLOP. Scores and ranks in bold font indicate ligands that rank in the top 200 for the respective scoring function.
Concentration at which ligand was tested.
ΔTm monitored using fluorescence, exciting at λ = 283 nm and measuring the integrated emission above 300 nm.
NR is not ranked.
919
920
that do not easily fit into the unminimized cavity
into which they were docked, explaining their poor
docking ranks, but they fit well upon receptor relaxation by MM–GBSA. Binding was detected at
millimolar concentrations by temperature of melting
(Tm) upshift experiments for seven of these nine
compounds; however, for two no binding was detected (Table 1). AMBERDOCK correctly predicted
binding for five ligands and incorrectly predicted
binding for 1-phenylsemicarbazide (3) and 2-phenoxyethanol (9) (two of the prospectively tested molecules were not rescored by AMBERDOCK because docking ranked them worse than 5000). PLOP
correctly predicted binding for five ligands, while
Rescoring Docking Hit Lists
incorrectly predicting binding for 2-phenoxyethanol
(9). PLOP agreed with docking on the remaining
three molecules that had been prioritized by AMBERDOCK, ranking them worse than 1000. Two of these,
4-(methylthio)nitrobenzene (2) and 2-ethoxyphenol
(5), were true ligands and so are false negatives for
PLOP.
Five high-resolution (better than 2 Å) protein–
ligand crystal structures were obtained for these
new L99A ligands to compare experimental to predicted poses (Fig. 3). In each case, electron density
for the ligands was unambiguous, allowing us
to model their positions in the site. Docking and
MM–GBSA methods predicted the binding
Fig. 3. Predicted and experimental ligand orientations for the hydrophobic L99A cavity. The carbon atoms of the
crystallographic pose, the DOCK predicted pose, the AMBERDOCK predicted pose, and the PLOP predicted pose are
colored gray, yellow, cyan, and magenta, respectively. The Fo − Fc omit electron density maps (green mesh) are contoured
at 2.5–3.0σ (a) β-chlorophenetole (1), (b) 4-(methylthio)nitrobenzene (2), (c) 2,6-difluorobenzylbromide (4), (d) 2ethoxyphenol (5), and (e) 3-methylbenzylazide (6) bound to L99A. Rendered with the program PyMOL.26
Rescoring Docking Hit Lists
1.93/1.84a
NA
1.80/1.70a
1.16
0.93
1.02
1.44
0.87
1.61
0.97
NA
0.63
2.04/1.12a
2.00/0.81a
2.00/0.65a
All crystals belong to space group P3221. Values in parentheses are for the highest resolution shell.
NA, not available.
a
Two conformations of the crystallographic ligand were modeled.
1.29
0.91
1.10
1.42
0.83
1.08
0.62
0.28
0.37
1.44
1.46
1.49
0.82
0.64
0.54
0.60
0.52
0.58
1.56 (1.60)
28,172 (1930)
10.6 (36.1)
99.3 (93.6)
14.6 (5.0)
18.3 (19.6)
20.5 (25.6)
0.01
1.13
2RBS
1.43 (1.47)
35,537 (2423)
7.6 (37.6)
99.4 (93.1)
17.3 (4.2)
18.1 (23.4)
20.1 (27.1)
0.01
1.13
2RBR
1.63 (1.68)
24,034 (1662)
6.4 (45.1)
97.0 (93.1)
14.5 (2.9)
20.8 (41.4)
24.0 (55.0)
0.02
1.58
2RBQ
1.47 (1.51)
33,813 (2446)
7.1 (46.6)
99.8 (99.6)
24.3 (4.5)
18.1 (24.6)
20.8 (28.6)
0.01
1.10
2RBP
1.29 (1.32)
48,915 (3563)
6.7 (62.2)
99.9 (100.0)
26.9 (3.3)
17.2 (23.5)
19.1 (28.6)
0.01
1.22
2RBO
1.29 (1.32)
48,474 (3576)
8.0 (56.9)
98.8 (100.0)
21.0 (3.6)
17.8 (21.4)
19.1 (21.6)
0.01
1.07
2RBN
1.46 (1.50)
33,797 (2337)
9.3 (34.8)
99.5 (94.9)
13.8 (3.6)
18.0 (25.6)
21.3 (32.3)
0.01
1.08
2RB2
1.70 (1.74)
21,923 (1615)
7.4 (63.2)
99.2 (99.9)
15.2 (2.4)
19.1 (32.4)
23.0 (38.9)
0.01
1.23
2RB1
1.84 (1.89)
13,875 (772)
7.7 (38.9)
80.0 (60.6)
11.8 (2.6)
19.6 (32.3)
23.3 (34.3)
0.01
0.99
2RB0
1.64 (1.68)
24,246 (1780)
6.1 (45.4)
99.8 (99.9)
22.8 (3.3)
19.1 (34.9)
22.1 (44.3)
0.01
1.15
2RAZ
1.80 (1.84)
18,414 (1314)
7.0 (50.0)
99.7 (98.9)
23.2 (3.4)
18.7 (27.8)
21.2 (30.2)
0.01
1.25
2RAY
Resolution (Å)
Reflections
Rmerge (%)
Completeness (%)
⟨I⟩/⟨σ(I)⟩
R-factor (%)
Rfree (%)
Δbond lengths (Å)
Δbond angles (°)
PDB code
RMSD (Å)
DOCK
AMBER
PLOP
3-Chloro-1phenyl1-propanol (13)
2Phenoxy
ethanol (9)
3-Methyl
benzyl
azide (6)
2-(n-Propyl thio)
ethanol (12)
n-Phenyl
glycinonitrile
(10)
2-Nitrothiophene
(11)
L99A/M102Q ligands (ID)
3-Methyl
benzyl
azide (6)
2-Ethoxyphenol
(5)
2,6-Difluoro
benzyl
bromide (4)
4-(Methylthio)
nitrobenzene
(2)
β-Chloro
phenetole
(1)
Ten representative compounds that scored well by
the MM–GBSA methods were experimentally tested
for binding to the polar cavity (Table 3). These compounds were ranked poorly by docking, again typically because they were too large for the conformation of the cavity targeted by docking. Binding was
detected at millimolar concentrations by Tm upshift
for 6 of these 10 compounds; for the remaining 4,
binding was not observed (Table 3). We note, however, that for 1 of these 4, 2-(n-propylthio) ethanol
(12), we were able to determine a crystal structure
in complex with the ligand by soaking a crystal
of L99A/M102Q with 100 mM of compound, suggesting that it is a weak ligand for this cavity.
AMBERDOCK correctly predicted binding for 4 of
the 6 ligands that it suggested should bind, while
incorrectly predicted binding for o-benzylhydroxylamine (14) and 1-phenylsemicarbazide (3). Of
the remaining two hits tested, prioritized by a high
PLOP ranking, AMBERDOCK missed 1 real ligand
but correctly distinguished 1 real decoy, ranking
both compounds worse than 500. Two of the prospectively tested molecules were not rescored by
AMBERDOCK because docking ranked them worse
L99A ligands (ID)
New L99A/M102Q ligands predicted by
rescoring
Table 2. Crystallographic measurement and the RMSD values for predicted and crystallographic ligand geometries in the L99A and L99A/M102Q sites
geometry for three of the five ligands to within 0.3
to 0.8 Å RMSD (Table 2). Conversely, the docked
pose of 3-methylbenzylazide (6) was 1.4 Å RMSD
from the crystallographic pose. The PLOP minimized prediction had a slightly improved RMSD of
1.1 Å, but the refined ligand also had a nonlinear
azide group, highlighting a failure in ligand
parameterization. In addition, docking and MM–
GBSA methods predicted poses that were approximately 1.5 Å RMSD from the crystallographic pose
of 4-(methylthio)nitrobenzene (2). The crystallographic poses of these two ligands would have been
within 2 Å of the Val111 side chain in the conformation of the cavity used for the docking
calculation, a steric conflict that is relieved by conformational expansion of the cavity in the experimental structures. Indeed, for all complexes, with
the exception of β-chlorophenetole (1), the F-helix
of lysozyme (residues 108–113) that forms one wall
of the cavity reorients by about 2 Å and swings
Val111 further out of the cavity to accommodate
the ligands.40 The protein conformations seen in
these structures more closely resemble the larger
isobutylbenzene-bound cavity site [Protein Data
Bank (PDB) ID 184L] than the smaller benzenebound cavity site (PDB ID 181L) used for docking
and rescoring. Whereas the MM–GBSA methods
do not capture this helix motion, receptor and
ligand minimization reduces the steric clash sufficiently to improve the ranks of what were docking false negatives. Higher level calculations
using free-energy methods and MD have captured
the F-helix motion and explained discrepancies
in free energies upon ligand binding due to its
displacement.21,31
921
922
Table 3. Compounds predicted by AMBERDOCK and PLOP to bind to T4 lysozyme L99A/M102Q
Score and ranka
Structure
Compound (ID)
DOCK
O-Benzylhydroxylamine (14)
− 11.35 (647)
− 28.05 (1)
−14.14 (2271)
10
6.8
10
6.8
Binding
detected
Structure
determined
No
No
0.0c
No
No
ΔTm (°C)
− 0.6
−16.42 (1354)
N-Phenylglycinonitrile (10)
− 8.60 (1556)
− 25.47 (11)
−40.17 (11)*
b10
3.0
5.1
Yes
Yes
− 24.52 (13)
−16.94 (1165)
b10
3.0
4.4
Yes
Yes
− 24.21 (14)
−15.18 (1824)
10
3.0, 6.8
Weak
No
NRd
−27.20 (20)
10
3.0
0.1
No
Yes
− 12.82 (318)
− 7.14 (2215)
6.02 (6847)
1.3, − 0.8
− 1.58 (4291)
− 10.25 (2260)
−27.19 (21)
10
3.0
0.0
No
No
3-Methyl benzyl azide (6)
− 5.35 (2740)
− 20.51 (116)
−25.87 (35)
10
3.0
1.9
Yes
Yes
2-Phenoxyethanol (9)
− 4.08 (3270)
− 16.53 (551)
−25.82 (36)
10
3.0
1.2
Yes
Yes
NRd
−25.65 (37)
10
3.0
7.8
Yes
Yes
3.6 (6074)
Compound scores and ranks (in parenthesis) for DOCK, AMBERDOCK, and PLOP. Scores and ranks in bold font indicate ligands that rank in the top 200 for the respective scoring function.
Concentration at which ligand was tested.
ΔTm monitored using fluorescence at λ = 291.5 nm and measuring the integrated emission above 300 nm.
NR, not ranked.
Rescoring Docking Hit Lists
cis-2-Hexen-1-ol (16)
(R) (+)-3-Chloro-1-phenyl-1-propanol (13)
d
pH
− 26.79 (4)
2-(n-Propyl-thio)ethanol (12)
c
(mM)
− 3.76 (3783)
2-Ethoxy-3,4-dihydro-2H-pyran (15)
b
PLOP
AMBER
1-Phenylsemi-carbazide (3)
2-Nitrothiophene (11)
a
|C|b
Rescoring Docking Hit Lists
than 5000. PLOP correctly predicted binding for 5 of
the 6 ligands that it suggested should bind but
incorrectly predicted binding for cis-2-hexenol (16).
Of the remaining hits tested, prioritized for testing
by AMBERDOCK, PLOP missed 2 true ligands but
correctly distinguished 2 decoys by ranking them
worse than 1000.
Crystal structures of six L99A/M102Q ligand
complexes were determined to compare predicted
and experimental poses of these new ligands (Fig. 4).
Electron density for each ligand was unambiguous
and was detailed enough to suggest two binding
modes for 2-nitrothiophene (11) and 3-chloro-1-
923
phenyl-1-propanol (13). Docking predicted the pose
of one ligand, 2-(n-propylthio)ethanol (12), to within
1 Å RMSD, while AMBERDOCK further minimized
five of its six ligands and PLOP minimized three of
its six ligands to within 1 Å RMSD (Table 2). Although the MM–GBSA methods collectively improved the binding mode predictions of all but one
ligand, the key hydrogen bond interaction was
missed in three of these structures (Fig. 4a, e, and f).
In addition, the azide group of 3-methylbenzylazide
(6) was incorrectly parameterized by both AMBERDOCK and PLOP, as was also observed in the L99A
cavity. Neither DOCK nor the MM–GBSA rescoring
Fig. 4. Predicted and experimental ligand orientations for the polar L99A/M102Q cavity site. The carbon atoms of the
crystallographic, DOCK, AMBERDOCK, and PLOP predicted poses are colored gray, yellow, cyan, and magenta,
respectively. Hydrogen bonds are depicted with dashed lines. The Fo − Fc electron density omit maps (green mesh) are
contoured at 2.5–3.0σ. (a) n-Phenylglycinonitrile (10), (b) 2-nitrothiophene (11), (c) 2-(n-propylthio)ethanol (12), (d)
3-methylbenzylazide (6), (e) 2-phenoxyethanol (9), and (f) (R)-(+)-3-chloro-1-phenyl-1-propanol (13) bound to L99A/
M102Q. Rendered with the program PyMOL.26
924
Table 4. Compounds predicted to bind by AMBERDOCK and PLOP to CCP W191G
Score and ranka
Structure
|C|b
(mM)
Structure
determined
10.0
No
No
− 16.21 (952)
1.0
Yes
Yes
347.86 (39)
− 44.39 (389)
1.0
Yes
Yes
− 14.74 (1830)
348.17 (49)
− 31.88 (796)
0.05
Yes
Yes
5-Nitro-6-aminouracil (21)
− 12.14 (2435)
348.49 (62)
− 31.47 (827)
1.0
No
No
1,2-Dimethyl-1H-pyridin-5-amine (22)
− 22.95 (362)
349.34 (87)
− 54.67 (59)
0.5
Yes
Yes
2-Aminobenzylamine (23)
− 12.62 (2316)
349.34 (96)
− 34.19 (671)
10.0
No
No
− 36.54 (7)
344.29 (12)
− 59.87 (53)
1.0
Yes
Yes
DOCK
N-Methyl-1,2-phenylenediamine (17)
− 20.6 (618)
347.03 (30)
− 38.08 (530)
N-Methylbenzylamine (18)
− 18.59 (942)
347.85 (38)
Cyclopentane carboximidamide (19)
− 13.38 (2134)
(1-Methyl-1H-pyrrol-2-yl)-methylamine (20)
Pyrimidine-2,4,6-triamine (24)
AMBER
PLOP
Rescoring Docking Hit Lists
Binding
detected
Compound (ID)
a
b
− 8.52 (3093)
363.47 (1901)
−56.65 (32)
10.0
No
No
1-Methyl-5-imidazolecarboxaldehyde (26)
− 21.14 (551)
358.53 (746)
−57.12 (28)
10.0
Yes
Yes
3-Methoxypyridine (27)
− 23.05 (2665)
355.17 (393)
−55.31 (44)
10.0
Yes
Yes
2-Imino-4-methylpiperidine (28)
− 17.30 (1695)
349.17 (82)
−52.43 (119)
10.0
Yes
Yes
2,4,5-Trimethyl-3-oxazoline (29)
− 13.96 (1962)
355.98 (455)
−52.32 (124)
0.25
Yes
Yes
1-Methyl-2-vinylpyridinium (30)
− 15.17 (1716)
363.60 (1938)
−52.32 (125)
0.5
Yes
Yes
Rescoring Docking Hit Lists
1,3-Dimethyl-2-oxo-2,3-dihydro-pyrimidin-1-ium (25)
Compound scores and ranks (in parenthesis) for DOCK, AMBERDOCK, and PLOP. Scores and ranks in bold font indicate ligands that rank in the top 200 for the respective scoring function.
Concentration at which ligand was tested.
925
926
correctly predicted the binding mode for 3-chloro-1phenyl-1-propanol (13), with RMSD values of 1.9 and
1.7 Å, respectively. In three structures—2-nitrothiophene (11), 3-methylbenzylazide (6), and 3-chloro-1phenyl-1-propanol (13)—the F-helix of the cavity
moves to accommodate the ligands while keeping
the cavity still buried from solvent. In the complexes
with 2-(n-propylthio)ethanol (12) and 2-phenoxyethanol (9), there is evidence of a second conformation of
residue Phe114 within the cavity that rotates and
opens a water channel to the surface of the protein.
Neither the helix movement nor the Phe114 rotation
was sampled by the MM–GBSA methods.
New W191G ligands predicted by rescoring
Fourteen representative compounds reprioritized
to score well by the MM–GBSA methods but scored
poorly by docking were experimentally tested for
binding by measuring perturbation of the heme
Soret band in CCP (Table 4).24 Binding was detected for 10 of these compounds at concentrations
ranging from 50 μM to 10 mM. Of the 11 compounds that AMBERDOCK predicted to bind with
ranks better than 500, binding was detected for 8.
Of the remaining prospective hits tested, AMBERDOCK correctly distinguished 1 compound as a
decoy but missed 2 ligands by ranking them worse
than 500. Of the 9 compounds that PLOP predicted to bind with ranks better than 500, binding
was detected for 8. Of the remaining prospective
hits tested, PLOP missed 2 ligands but correctly
distinguished 3 decoys, ranking them worse than
500.
Crystal structures of CCP in complex with the 10
new ligands were obtained (Fig. 5). The electron
density for the ligands was unambiguous. Docking
predicted three structures to within 1 Å of the crystallographic result, whereas the MM–GBSA methods
did so for 7 structures, typically with improved hydrogen-bonding interactions (Table 5). For 3 ligands,
the docking poses were over 1.9 Å away from the
crystallographic results, and MM–GBSA refinement
did little to improve these structures. In 4 of the
complex structures—cyclopentane carboximidamide
(19), 1,2-dimethyl-1H-pyridine-5-amine (22), pyrimidine-2,4,6-triamine (24), and 1-methyl-2-vinylpyridinium (30)—the loop composed of residues 190–195
flips out by nearly 12 Å, opening the cavity to bulk
solvent. This large loop motion was not sampled by
MM–GBSA.
Overall performance in predicting top 100 hits
The simplicity of these model cavity sites, the
number of known ligands and decoys, and our
Rescoring Docking Hit Lists
experience with their ligands21,23,28–30 often allow us
to predict what turn out to be true ligands and true
decoys from among top-scoring molecules, based on
their physical properties. We examined the top 100
hits predicted to bind by docking and MM–GBSA,
compared property distributions, and made educated guesses as to whether or not they will bind. The
100 top-ranking MM–GBSA rescored compounds
for the L99A and L99A/M102Q cavities were larger,
more flexible, and more polar, with more hydrogenbond acceptors and lower ClogP values per heavy
atom compared to the top 100 hits from docking. For
the anionic W191G cavity there was a similar trend
towards larger molecules and also a drift away from
the singly charged cations favored by DOCK, with
more dications and neutral molecules prioritized
among the top-ranking 100 molecules by the MM–
GBSA methods. The increased size and greater
differences in polarity of the molecules in the MM–
GBSA hit lists resulted in lower mean pairwise
similarities among the molecules and, consequently,
an increase in the diversity of the rescored hit lists
relative to the docking hit lists. Thus, using ECFP_4
fingerprints (SciTegic, Inc.), we found the average
pairwise Tanimoto coefficient among the 100 topdocking molecules for the L99A cavity with DOCK,
AMBERDOCK, and PLOP to be 0.17, 0.12, and 0.10,
respectively (full distributions of pairwise similarities are given in Supplementary Fig. S1). Similar
trends were observed in the other two cavities. The
same tendencies that led to greater diversity in
ligands and their properties, however, reduced the
raw hit rates we anticipate from among the top 100
ranking MM–GBSA ligands compared to those predicted by docking (Table 6). For example, among the
top 100 docking hits for the CCP cavity there were 29
true ligands and no experimentally determined
decoys. Of the remaining molecules—all untested—
were what we predict to be 79 likely ligands and 7
likely decoys, based on their similarity to known
ligands and decoys and their physical properties such
as size and charge complementarity. Conversely,
among the top 100 PLOP hits for the anionic cavity
were only 15 experimentally tested ligands and 1
experimental decoy. Among the untested molecules
were what we suspect are 53 further ligands and 22
further decoys. Among the top AMBERDOCK hits
for this cavity were 19 true ligands and 3 experimental decoys. Among the untested molecules
prioritized by this program, we suspect that there
are 67 further ligands and 14 more decoys. Similar
trends were observed in the other two cavities (Table
6). Admittedly, these numbers reflect guesses only,
but we suspect that the overall trends would be born
out by experiment (the interested reader may draw
their own conclusions from the full lists in Supple-
Fig. 5. Predicted and experimental ligand orientations for the anionic CCP cavity. The carbon atoms of the crystallographic, DOCK, AMBERDOCK, and PLOP predicted poses are colored gray, green, cyan, and orange, respectively.
(a) n-Methylbenzylamine (18), (b) cyclopentane carboximidamide (19), (c) (1-methyl-1H-pyrrol-2-yl)-methylamine (20),
(d) 1,2-dimethyl-1H-pyridin-5-amine (22), (e) pyrimidine-2,4,6-triamine (24), (f) 1-methyl-5-imidazolecarboxaldehyde
(26), (g) 3-methoxypyridine (27), (h) 2-imino-4-methylpiperdine (28), (i) 2,4,5-trimethyl-3-oxazoline (29), and (j) 1-methyl2-vinylpyridinium (30). Rendered with the program PyMOL.26
Rescoring Docking Hit Lists
927
Fig. 5 (legend on previous page)
Rescoring Docking Hit Lists
2.58
2.55
2.59
0.55
0.50
0.33
1.32
0.83
0.81
L99A cavity
DOCK
PLOP
AMBERDOCK
All crystals belong to space group P212121. Values in parentheses are for the highest resolution shell.
a
Two conformations of the crystallographic ligand were modeled.
1.07
1.06
0.94
1.91
1.70
1.91
2.97/3.16a
2.84/3.01a
2.98/3.15a
0.58
0.34
0.30
1.34
0.31
0.99
1.05
0.37
0.45
Likely Likely
True
True ligands decoys
ligands decoys in top in top
in top in top
100
100
hits
hitsb Ambiguousc
hits
hitsa
7
6
8
3
1
2
63
35
54
23
22
25
14
43
21
L99A/M102Q cavity
DOCK
13
PLOP
5
AMBERDOCK
7
2
1
2
73
31
43
12
8
22
15
61
35
W191G cavity
DOCK
PLOP
AMBERDOCK
0
1
3
79
53
67
7
22
14
14
25
19
29
15
19
a
Molecules that, based on their physical properties and similarity to known ligands, are likely to be cavity ligands (a full list is
given in Supplementary Tables S1–S9).
b
Molecules that, based on their physical properties and
similarity to known decoys, are likely not to bind.
c
Molecules that are sufficiently different from known ligands
and decoys and whose physical properties are not sufficiently
distinctive, such that no prediction was made (for L99A and L99A/
M102Q molecules). For W191G, molecules that were misprotonated during database preparation relative to the expected protonation at pH 4.5 are not counted to measure the performance of the
scoring function.
0.89
0.81
0.77
1.50 (1.54)
63,108 (3364)
4.2 (19.6)
99.3 (95.4)
39.2 (5.9)
14.6 (16.3)
16.7 (21.7)
0.01
1.14
2RC2
2.49 (2.56)
11042 (680)
2.7 (5.9)
82.3 (69.7)
26.1 (12.9)
17.7 (21.9)
22.9 (29.4)
0.02
1.95
2RC1
1.50 (1.54)
63,783 (4467)
5.0 (33.4)
99.7 (96.5)
31.0 (3.5)
14.3 (17.9)
16.9 (23.9)
0.01
1.22
2RC0
1.50 (1.54)
63,445 (4489)
3.8 (22.1)
99.7 (96.3)
37.8 (5.2)
14.5 (16.9)
16.4 (21.0)
0.01
1.24
2RBY
1.50 (1.54)
63,009 (4512)
5.2 (38.5)
99.8 (97.9)
30.6 (2.8)
15.2 (21.0)
17.3 (25.7)
0.01
1.14
2RBX
1.80 (1.85)
36,504 (2670)
5.3 (10.9)
99.4 (99.9)
51.7 (30.1)
15.9 (19.3)
19.5 (24.0)
0.02
1.39
2RBU
1.39 (1.43)
78,561 (5188)
4.5 (20.2)
99.0 (90.5)
43.2 (5.6)
13.8 (20.7)
15.4 (24.1)
0.01
1.19
2RBV
1.50 (1.54)
57,782 (4371)
5.1 (22.4)
92.0 (95.9)
41.7 (6.0)
14.4 (17.8)
16.7 (25.5)
0.01
1.22
2RBW
1.80 (1.85)
36,874 (2587)
6.4 (23.3)
99.5 (96.2)
34.8 (9.2)
15.0 (18.4)
19.1 (24.0)
0.02
1.50
2RBZ
Method
1.24 (1.27)
104,081 (4744)
6.1 (40.7)
93.1 (58.0)
29.9 (1.8)
12.2 (24.8)
14.6 (24.4)
0.01
1.59
2RBT
n-Methylbenzylamine
(18)
Table 6. Likely ligands and decoys among the top 100
ranked ligands by docking and MM–GBSA
Resolution (Å)
Reflections
Rmerge (%)
Completeness (%)
⟨I⟩/⟨σ(I)⟩
R-factor (%)
Rfree (%)
Δbond lengths (Å)
Δbond angles (°)
PDB code
RMSD (Å)
DOCK
AMBER
PLOP
1-Methyl-2vinylpyridinium
(30)
1-Methyl-52,4,5imidazole
2-Imino-4carboxaldehyde 3-Methoxypyridine methylpiperidine Trimethyl-3oxazoline (29)
(26)
(27)
(28)
1,2-Dimethyl- Pyrimidine1H-pyridin-52,4,6amine (22) triamine (24)
(1-Methyl-1Hpyrrol-2-yl)methylamine
(20)
Cyclopentane
carbox
imidamide
(19)
Table 5. Crystallographic measurement and the RMSD values for predicted and crystallographic ligand geometries in the CCP site (compound IDs in parentheses)
928
mentary Tables 1–9) Thus, whereas the MM–GBSA
methods rescued many docking false negatives and
sampled a more diverse chemical space among the
top hits, they also suggested more false positives
among the very top scoring molecules and, we
suspect, have a lower overall hit rate in this segment
of the molecules prioritized for testing.
Origins of false-positive hits suggested by
MM–GBSA rescoring
In these simple cavities, false-positive hits often
identify specific pathologies in a scoring function.
For example, the MM–GBSA methods seemed
distracted by compounds bearing what is almost
certainly the wrong net charge for the W191G cavity,
which extensive testing has shown preferentially
binds monocations over neutral molecules (few of
which have been observed to bind, and then only
weakly) and dications (none of which have been
observed to bind). For instance, among the top 100
ranking molecules predicted by PLOP, there were 13
dications. Whereas AMBERDOCK predicted only
one dication, it prioritized five neutral molecules
among the top 100 hits. The dications will pay too
high a desolvation penalty to be compensated for by
the interaction with the single anion in the site
(Asp235), and the neutral compounds desolvate the
same aspartate without recouping enough in interaction energy. Balancing polar and ionic interactions
with concomitant solvation penalties is a challenge
for the field, one clearly faced by these methods as
Rescoring Docking Hit Lists
well. On the other hand, many of the top-ranked
PLOP ligands for L99A (47 out of the top 50) and
L99A/M102Q contained one or more nitriles. While
some of these compounds may well be ligands, as in
the case of n-phenylglycinonitrile (10) for L99A/
M102Q (Table 3), we suspect that this represents a
ligand parameterization problem as opposed to a
genuinely meaningful enrichment. Indeed, the PLOP
solvation energies for the top 47 nitriles actually
rewarded desolvation, rather than penalizing it as is
almost always the case otherwise, suggesting that
there is an issue with determining the correct selfenergy for this functional group. Not wishing this
term to dominate our analysis, we excluded these
compounds from the PLOP rescored hit lists for
L99A and L99A/M102Q; they do not contribute to
the accounting described in this work. This highlights the importance of good ligand parameterization for database screening—which is a considerable
challenge for hundreds of thousands of molecules
typically screened by docking—the lack of which can
undermine any improvement in theory.
Discussion
In principle, the most important improvements of
MM–GBSA over docking, certainly over the program used in this study, DOCK3.5.54, are the better
representation of electrostatic interactions, ligand
and protein desolvation energies, and relaxation of
the ligand–protein complex. The simplicity of the
model cavity sites allows us to explore how these
terms influence docking results in detail and to
make prospective predictions for ligands that we
can, in fact, acquire and test. Many investigators will
be unsurprised to see that the MM–GBSA methods
can rescue molecules that rank poorly in the docking
calculation owing to the rigid-receptor approximation used in docking. Ligands that were too big to be
accommodated well in the original docking are well
fit by a binding site that has been allowed to relax by
energy minimization and, in the case of AMBERDOCK, short MD simulations. This was true both in
retrospective calculations as well as in prospective
predictions. The ability to relax the site also resulted
in rescored hit lists that were more diverse with a
wider range of likely ligands. Perhaps less anticipated was the cost of allowing such conformational
change—some of the rescued, high-scoring molecules by MM–GBSA do not, in fact, bind to the
cavity sites. These molecules are new false positives
introduced by the higher level of theory. Indeed, the
overall hit rates at the very top of the ranked lists are
arguably better by simple docking than by MM–
GBSA rescoring, at least when evaluated simplistically by the raw number of hits and likely hits (this is
arguably offset by the greater diversity of the MM–
GBSA hit lists). Partly this reflects problems in
ligand parameterization, and partly difficulties in
the treatment of the electrostatics in the binding
sites. The most important challenge for MM–GBSA
and for flexible receptor models in general is
929
balancing the opportunities to find new ligands as
receptor geometries are relaxed with the introduction of new false positives as the need to consider
large receptor internal energies is introduced.
Specific examples of these opportunities and problems are apparent in the three cavity sites studied
here.
The principal improvement conferred by MM–
GBSA rescoring in the model cavity sites over
docking was the inclusion of receptor binding site
relaxation, which improved the ranks of larger
ligands that rigid receptor docking missed. AMBERDOCK, for example, correctly predicted 2-ethoxyphenol (5) to bind to L99A (Table 1, Fig. 3d). This
compound is too large for the unrelaxed conformation of this cavity targeted by docking, but minimization and MD simulations allow the ligand to be
well accommodated by effectively expanding the
site. Often, this relaxation led not only to improved
rankings but also improved geometries. For many
ligands, RMSD values between the MM–GBSA
predictions and the crystallographic results declined
relative to those of the docking predictions and,
especially in the W191G anionic cavity, many
ligands refined by MM–GBSA had improved hydrogen bonding to the site. Examples of this include
the new W191G cavity ligands n-methylbenzylamine (18) and cyclopentane carboximidamide (19)
(Table 4, Fig. 5a and b, respectively).
The structural relaxation with MM–GBSA performed well when the initial docking geometry
resembled the crystallographic pose, but did little
when large protein conformational changes were
provoked by ligand binding. For instance, F-helix
unwinding and rotamer change by Val111 in L99A
and L99A/M102Q were never captured by the
method, nor was the extensive loop flipping observed in several of the W191G–ligand complexes.
When such movements occurred, MM–GBSA
rescoring could not rescue substantially incorrect
docking poses, such as that adopted by 3-chloro-1phenyl-propanol (13) for L99A/M102Q (Table 3,
Fig. 4f) and pyrimidine-2,4,6-triamine (24) predicted
for CCP (Table 4, Fig. 5e), notwithstanding the large
improvement in their rankings conferred by the
rescoring. These large movements are outside the
radius of convergence of the local relaxation undertaken by the MM–GBSA methods. Indeed, even
more time-consuming thermodynamic integration
methods are hard put to sample such changes without explicit “confine-and-release” strategies, which
depend on a foreknowledge that such movements
are likely.41 And whereas loop sampling methods
have had encouraging successes in predicting such
large movements,42 this remains a frontier challenge
for ligand and structure prediction methods.
Pragmatically, the inability to predict the structural accommodations provoked by some large
ligands is offset by the correct reprioritization of
what were docking false negatives as ligands. The
same comfort is not afforded by the 10 false negatives
introduced by the MM–GBSA methods, nor by the
lower overall hit rates compared to docking among
930
the very top scoring ligands (Table 6). By allowing
the receptor to respond to ligand binding, one allows
for new and potentially unfavorable receptor conformations. These must be distinguished by the
MM–GBSA energy functions from the true lowenergy conformations that may be sampled in
solution. This is challenging, as the receptor conformational energies are large, and the errors in these
calculations are typically on the same order of the net
interaction energy of the protein–ligand complex.
Although some of the errors are cancelled by
subtraction of the internal energies before and after
ligand binding, one is still subtracting two large
numbers with relatively large errors to find a small
one, the net binding free energy. Consistent with this
view, ligands achieved their maximal advantage
over decoys on rescoring when we allowed only a
5 Å region around the binding site to relax. Allowing
the full protein to relax, or even an 8 Å region around
the binding site, diminished the discrimination of
known ligands from decoys. Of course, relaxing the
entire system is the more physically correct way to
calculate these energies. Falling back on limited
relaxation speaks of a larger methodological issue.
The three cavity sites targeted here are contrivances of human design, and ligands discovered for
them have no intrinsic value other than for testing
methods. Indeed, in these simple model systems the
failures are often more interesting than the successes,
as they can illuminate a specific methodological
problem.21,23,28–30 Examples are the 10 false positives
predicted for the cavity sites by MM–GBSA rescoring. Some of these reflect ligand parameterization
problems. For instance, we suspect that the many
nitrile-containing decoys predicted by PLOP for
L99A and L99A/M102Q reflect failures in ligand
parameterization. Such mechanical failures may be
addressed by close attention to particular ligand
groups and improved partial atomic charge models;
admittedly, this can be a daunting task for screening
databases containing hundreds of thousands of
disparate molecules. More interesting are the 8
false positives that are true energy function decoys.
Several of these highlight difficulties in the treatment
of electrostatics and solvation in the binding sites. 2Phenyoxyethanol (9), for example, was predicted to
bind by both PLOP and AMBERDOCK to L99A
(Table 1). This decoy has a similar topology to 2phenylpropanol, a known ligand22 (Fig. 6a); however, the ether of 2-phenoxyethanol (9) increases its
polarity and presumably its solvation energy, which
is not fully captured by the MM–GBSA implicit
solvent model (another possibility would be that the
2-phenoxyethanol is docked in a high-energy conformation, one that is not recognized by the rescoring methods, but this turns out not to be the case,
with both the decoy and the ligand 2-phenylpropanol adopting similar and low-energy conformations). Similarly, o-benzylhydroxylamine (14) was
the top-ranking AMBERDOCK hit for L99A/
M102Q, but is a decoy (Table 3). The terminal
-ONH2 of this compound is too polar for the site,
stranding one unpaired polar hydrogen from the
Rescoring Docking Hit Lists
Fig. 6. The topologically similar ligands and decoys of
(a) 2-phenylpropanol and 2-phenoxyethanol (9) for L99A
and (b) n-phenylhydroxylamine and o-benzylhydroxylamine (14) to L99A/M102Q.
NH2 group in this largely hydrophobic site. Interestingly, the polar cavity does bind n-phenylhydroxylamine (unpublished data), which has the same
hydrogen bond accounting as o-benzylhydroxylamine (14) and topologically resembles it closely (Fig.
6b). The difference between these nearly identical
molecules is that in the former the two hydrogenbond donors from the ligand can both be accommodated by the carbonyl of the receptor glutamine,
whereas in the decoy both hydrogen-bond donors
originate from the same atom—the nitrogen of the obenzylhydroxylamine (14)—and only one can be
accommodated by the carbonyl oxygen.
The challenges of balancing ligand electrostatic
interaction energies and desolvation penalties were
also apparent in the anionic W191G cavity. Most
obvious were those molecules that did not bear the
correct monocationic charge state. The 13 molecules
that were doubly charged among the top-scoring
PLOP hits are almost certainly decoys, and this is
also the case for the AMBERDOCK false-positive 5nitro-6-aminouracil (21), which is neutral and
cannot make the ion–pair interaction with Asp235
(Table 4). More subtly, whereas 1,3-dimethyl-2-oxo2,3-dihydropyrimidin-1-ium (25) is charged, this
charge is shared between the two cyclic nitrogen
atoms and results in a compound with reduced
electrophilicity compared to a compound with a
localized charge. The AMBERDOCK false positives
n-methyl-1,2-phenylene-diamine (17) and 2-aminobenzylamine (23) (Table 4) most likely do not bind
because of steric clashes that inhibit optimal
positioning of the charge–charge interaction. These
failures point to specific directions for improved
treatment of the balance between electrostatic
interaction and desolvation energies in the MM–
GBSA methods.
Overall, the results of MM–GBSA rescoring of
docking hit lists on the model binding sites seem
conflicted. On the one hand, rescoring rescued many
docking false negatives, improved the geometric
fidelity of most of the predicted structures, and
increased the diversity of the hit lists. On the other
hand, rescoring introduced more false positives,
especially among the very top ranking ligands,
Rescoring Docking Hit Lists
compared to the simpler docking protocol. These
observations may be reconciled by recognizing
that what is probably the greatest advantage of the
MM–GBSA methods over docking for the model
sites, the relaxation of the protein–ligand complex,
also presents the greatest challenge to discrimination. To allow a flexible receptor, one must consider
the relative energies of the different protein conformations explored. This implicates the pairwise
interactions of thousands of protein atoms, as
opposed to the tens of atoms involved in the
immediate protein–ligand complex. To properly
rank the energies of the complexes, one must also
properly account for the larger uncertainties that
accompany the much higher magnitude energies of
the overall system. Whereas this is the thermodynamically correct approach, it introduces many
interactions that have little bearing on the intimacies
of the protein–ligand complex itself. Rigid receptor
docking, for all the calumny poured upon it, can
ignore these large-magnitude yet low-relevance
interactions. Of course, this leads to many false
negatives, but it avoids many of the false positives to
which the MM–GBSA methods are prone. Pragmatically, this suggests that hits derived from docking
to a rigid experimental receptor conformation—and
ideally more than one30,43—and hits prioritized by
rescoring after MM–GBSA refinement with binding
site minimization will provide good candidates for
experimental testing. Despite its greater sophistication, MM–GBSA rescoring has a harder task, and its
predictions will not, by every criterion, be better
than those of a modern docking program; rather, our
results suggest they will complement and add to
them. Still, MM–GBSA is a higher level of theory,
and because it is grounded in physics, they can be
built upon and improved in a regular way. They are
thus on a path to fundamental improvement in
molecular docking and structure-based screening,
which is so actively sought.44
931
Rescoring with PLOP
The rescoring procedure with PLOP36,37 was essentially
as described.17 Ligand parameters were calculated with
IMPACT.47 The partial atomic charges of the ligands were
replaced by the AM1-CM2 charges calculated by AMSOL
(v6.5.3) as these were the same charges used during the
initial docking.23 The same protein structure file used in
docking was used for rescoring. Protein parameters were
defined by IMPACT with the exception of the partial
charges for the heme cofactor in CCP W191G, which were
the same as used in the docking method.28 All energy
minimizations were performed using PLOP with the allatom OPLS force field (OPLS-AA)48 and the surface
generalized Born (SGB) implicit solvent model.49 PLOP
implements a multiscale truncated-Newton minimization
algorithm as described.50 For receptor minimization and
calculation of Ecomplex and Ereceptor, residues in a prespecified list within 5 Å of the binding site were minimized after an initial side-chain rotamer search. (Residues
78, 84, 85, 87, 88, 91, 98–100, 102, 103, 106, 111, 118, 121,
133, and 153 for L99A and L99A/M102Q and residues
174–180, 189–192, 202, 230–232, 235, and water 308 for
CCP). The rotamer search algorithm is as described in the
Supplementary Material.
Preliminary PLOP calculations of the hydrophobic and
polar cavities were performed with a rigid receptor and
resulted in very little separation of ligands and known
decoys. On the other hand, PLOP calculations in which a
larger set of residues (those within 8 Å of the binding site)
were minimized and resulted in worse overall enrichments of known ligands and a decreased separation of
known ligands and decoys relative to minimizing a
smaller 5 Å pocket. To approximate a fully desolvated
ligand and cavity for the hydrophobic L99A and polar
L99A/M102Q sites, only the SBG solvation term of the
free ligand was included in the calculation of the total
PLOP binding energies. Initial PLOP calculations including the SGB solvation terms for the calculation of the
complex and free protein energies resulted in poor enrichments of known ligands, decreased separation of
ligands and known decoys, as well as an enrichment of
hits with increased polarity and electrostatic interactions.
For the more solvated CCP cavity, the SGB terms were
included in the calculation of the complex, free protein,
and free ligand energies for the total binding energy.
Materials and Methods
Rescoring with AMBERDOCK
Docking against cavity sites
DOCK3.5.5423,38 was used to dock a multiconformer
database of small molecules into the model cavity sites.
The receptors, grids, spheres, and ligand databases were
prepared as described for the T4 Lysozyme23 and CCP28
cavities, respectively. Briefly, to sample ligand orientations, ligand, receptor, and overlap bins were set to 0.2 Å;
the distance tolerance for matching ligand atoms to
receptor was set to 0.75 Å. Each docking pose was evaluated for steric fit. Compounds passing this filter were
scored for electrostatic and van der Waals complementarity and assigned the full penalty for transfer from a
dielectric of 80 to a dielectric of 2, as calculated by
AMSOL.45,46 Sampling and scoring required less than a
second per ligand on a single 3.2-GHz Xeon processor. The
best scoring conformation of each of the 10,000 top-scoring
molecules against L99A and L99A/M102Q and the 5400
top-scoring molecules against CCP were saved and
rescored by the MM–GBSA protocols.
AMBERDOCK is based on the amber_score() scoring
module in DOCK6. The ligand structures were modified
using the antechamber suite of programs to create input files
that could be read by Leap to generate the parameter and
topology files for AMBERDOCK. Antechamber51 has been
developed to be used with the general AMBER force field
for small molecules.52 Charges for the ligands were
generated using three charge methods in AntechamberPEOE,53 AM1-BCC,54 and HF/6-31G* RESP.55 The protonation states of the ligands were kept the same as the
previous docking run for consistency in rescoring. AMBER
ff94 parameters were assigned to all the protein atoms. The
standard parameters for the heme cofactor as implemented
in the Amber 9 program was used for the CCP cavity.56 The
protonation states of histidine residues were predicted
based on their close neighbors. The GB model corresponding to igb = 5 in the AMBER 9 program was used.57 The
surface area term was calculated using the LCPO model.58
A nonbonded cutoff of 18 Å was used for the calculations.
Rescoring Docking Hit Lists
932
The starting structures were taken from the docked pose.
The structures were subjected to 100 steps of conjugate
gradient minimization, 3000 steps of MD simulation with a
1-fs time step at a temperature of 300 K, followed by 100
steps of minimization. During the minimization and MD,
only the ligand and the protein residues within 5 Å of the
ligand were allowed to move. To expedite the scoring
process, we calculated the energy of the receptor (Rreceptor)
once, and used this energy as a constant term during the
subsequent energy evaluations for the rest of the ligands in
the database. Binding-free-energy calculations with
AMBERDOCK follows a scheme as described in Supplementary Fig. S2. Several AMBERDOCK rescoring protocols with slight variations were retrospectively tested and
results are described in Supplementary Fig. S3.
Protein preparation and expression
T4 lysozyme mutants L99A and L99A/M102Q and CCP
mutant W191G were expressed and purified as described.23,24
Binding detection of ligands to T4 lysozyme cavities
by upshift of thermal denaturation temperature
To detect binding, L99A and L99A/M102Q were denatured reversibly by temperature in the presence and
absence of the putative ligand. Molecules that bind preferentially to the folded cavity-containing protein should
stabilize it relative to the apo protein, raising its temperature of melting.22 All thermal melts were conducted in
a Jasco J-715 spectropolarimeter as described.22 Each
compound was screened in its neutral form. All compounds tested against L99A and L99A/M102Q were
assayed in a pH 3 buffer containing 25 mM KCl, 2.9 mM
phosphoric acid, and 17 mM KH2PO4 with the exception
of 1-phenylsemicarbazide (3) and o-benzylhydroxylamine
(14). To maintain compound neutrality, these two were
assayed at pH 6.8 in a 50 mM potassium chloride and 38%
(v/v) ethylene glycol buffer.22 Thermal melts were monitored by far-UV circular dichroism, except for melts in the
presence of 4-(methylthio)nitrobenzene (2), 1-phenylsemicarbazide (3), and 2,6-difluorobenzylbromide (4), which
absorb strongly in the far-UV region. For these three,
thermal denaturation was measured by the intensity of the
integrated fluorescence emission for all wavelengths
above 300 nm, exciting at 283 to 292 nm, using a fluorescence detector on the Jasco instrument. Thermal melts
were performed at a temperature ramp rate of 2 K/min. A
least-squares fit of the two-state transition model was
performed with the program EXAM59 to calculate Tm and
van't Hoff ΔH values for the thermal denaturations. The
ΔCp was set to 8 KJ mol− 1 K− 1 (1.94 kcal mol− 1 K− 1).
Binding detection of ligands to CCP W191G
Ligand binding was measured in 50 mM acetate buffer,
pH 4.5. To avoid competition in ligand binding with small
cations such as potassium,24 the pH of the buffer was
adjusted with Bis–Tris propane. The compounds were
dissolved in dimethyl sulfoxide. Binding of compounds to
CCP was monitored by the red shift and increase of
absorbance of the heme Soret band24 at 10 °C.
crystallization buffer containing as much as 100 mM compound. In addition to soaking, drops of neat compound
were added to the cover slip surrounding the drop containing the crystal. After soaking, the crystals were cryoprotected with a 50:50 Paraton-N (Hampton Research,
Aliso Viejo, CA)/mineral oil mix. Crystals for CCP W191G
were grown as described25 and the resulting crystals belonged to space group P212121. Crystals were soaked in
25% methyl-2,4-pentanediol with 1 to 50 mM compound
for 4 h or overnight with the exception of pyrimidine2,4,6-triamine (24), which was soaked for 15 min.
Diffraction data for the complexes of L99A with β-chlorophenetole (1), 4-(methylthio)nitrobenzene (2), and 2,6-difluorobenzylbromide (4) and the complex of L99A/M102Q
with 3-methylbenzylazide (6) were collected using a Rigaku
X-ray generator equipped with a rotating copper anode and
a Raxis IV image plate. Data for the complexes of L99A/
M102Q with n-phenylglycinonitrile (10) and 2-nitrothiophene (11) and the complex of CCP with n-methylbenzylamine (18) were collected on beamline 9-1 at the Stanford
Synchrotron Radiation Laboratory using an ADSC CCD
detector. Data for all other complexes were collected on
beamline 8.3.1 of the Advanced Light Source at Lawrence
Berkeley National Laboratory using an ADSC CCD detector. All data sets were collected at 100 K. Reflections were
indexed, integrated, and scaled using HKL2000.60 Parameters for ligands were generated with PRODRG.61 Complexes were refined using the CCP4 software package.62
Interactive model building was performed using Coot.63
Supporting information available
A description of the PLOP side-chain rotamer search and
minimization algorithm, AMBERDOCK parameters and
optimization, and structures for the top 100 hits predicted
by DOCK, PLOP, and AMBERDOCK for the three cavity
sites is available online from the journal website.
Protein Data Bank accession codes
The crystallographic coordinates for the complex structures presented in this work have been deposited with the
RCSB Protein Data Bank with accession codes 2RAY,
2RAZ, 2RB0, 2RB1, 2RB2, 2RBN, 2RBO, 2RBP, 2RBQ,
2RBR, 2RBS, 2RBT, 2RBU, 2RBV, 2RBW, 2RBX, 2RBY,
2RBZ, 2RC0, 2RC1, and 2RC2.
Acknowledgements
This work was supported by GM59957 (to B.K.S.),
AI035707 (to M.P.J.) and GM56531 (to D.A.C.). M.P.J.
is a consultant to Schrodinger Inc. We thank Niu
Huang for advice on using PLOP; Michael Keiser
and Jerome Hert for help with chemical similarity
calculations; Michael Mysinger, Veena Thomas, and
Michael Keiser for reading this manuscript; and
MDL for providing the ACD database.
Structure determination
Supplementary Data
Crystals for L99A and L99A/M102Q were grown as
described23 and the resulting crystals belonged to space
group P3221. Crystals were soaked overnight to 1 week in
Supplementary data associated with this article
can be found, in the online version, at doi:10.1016/
j.jmb.2008.01.049
Rescoring Docking Hit Lists
933
References
1. Powers, R. A., Morandi, F. & Shoichet, B. K. (2002).
Structure-based discovery of a novel, noncovalent inhibitor of AmpC beta-lactamase. Structure, 10, 1013–1023.
2. Evers, A. & Klebe, G. (2004). Successful virtual screening for a submicromolar antagonist of the neurokinin1 receptor based on a ligand-supported homology
model. J. Med. Chem. 47, 5381–5392.
3. Grzybowski, B. A., Ishchenko, A. V., Kim, C.-Y.,
Topalov, G., Chapman, R., Christianson, D. W. et al.
(2002). Combinatorial computational method gives
new picomolar ligands for a known enzyme. Proc. Natl
Acad. Sci. USA, 99, 1270–1273.
4. Li, S., Gao, J., Satoh, T., Friedman, T. M., Edling, A. E.,
Koch, U. et al. (1997). A computer screening approach
to immunoglobulin superfamily structures and interactions: discovery of small non-peptidic CD4 inhibitors as novel immunotherapeutics. Proc. Natl Acad. Sci.
USA, 94, 73–78.
5. Huang, N., Nagarsekar, A., Xia, G., Hayashi, J. &
MacKerell, A. D., Jr (2004). Identification of nonphosphate-containing small molecular weight inhibitors of the tyrosine kinase p56 Lck SH2 domain via in
silico screening against the pY + 3 binding site. J. Med.
Chem. 47, 3502–3511.
6. Song, H., Wang, R., Wang, S. & Lin, J. (2005). A lowmolecular-weight compound discovered through virtual database screening inhibits Stat3 function in breast
cancer cells. Proc. Natl Acad. Sci. USA, 102, 4700–4705.
7. Lyne, P. D., Kenny, P. W., Cosgrove, D. A., Deng, C.,
Zabludoff, S., Wendoloski, J. J. & Ashwell, S. (2004).
Identification of compounds with nanomolar binding
affinity for checkpoint kinase-1 using knowledgebased virtual screening. J. Med. Chem. 47, 1962–1968.
8. Cavasotto, C. N., Oritz, M. A., Abagyan, R. A. &
Piedrafita, F. J. (2006). In silico identification of novel
EGFR inhibitors with antiproliferative activity against
cancer cells. Bioorg. Med. Chem. Lett. 16, 1969–1974.
9. Charifson, P. S., Corkery, J. J., Murcko, M. A. &
Walters, W. P. (1999). Consensus scoring: a method for
obtaining improved hit rates from docking databases
of three-dimensional structures into proteins. J. Med.
Chem. 42, 5100–5109.
10. Bissantz, C., Folkers, G. & Rognan, D. (2000). Proteinbased virtual screening of chemical databases. 1. Evaluation of different docking/scoring combinations.
J. Med. Chem. 43, 4759–4767.
11. Gohlke, H. & Klebe, G. (2001). Statistical potentials
and scoring functions applied to protein–ligand
binding. Curr. Opin. Struct. Biol. 11, 231–235.
12. Stahl, M. & Rarey, M. (2001). Detailed analysis of
scoring functions for virtual screening. J. Med. Chem.
44, 1035–1042.
13. Wang, R. X. & Wang, S. M. (2001). How does consensus scoring work for virtual library screening? An
idealized computer experiment. J. Chem. Inf. Comput.
Sci. 41, 1422–1426.
14. Friesner, R. A., Murphy, R. B., Repasky, M. P., Frye,
L. L., Greenwood, J. R., Halgren, T. A. et al. (2006).
Extra precision glide: docking and scoring incorporating a model of hydrophobic enclosure for protein–
ligand complexes. J. Med. Chem. 49, 6177–6196.
15. Wang, J., Kang, X., Kuntz, I. D. & Kollman, P. A.
(2005). Hierarchical database screenings for HIV-1
reverse transcriptase using a pharmacophore model,
rigid docking, solvation docking, and MM-PB/SA.
J. Med. Chem. 48, 2432–2444.
16. Kalyanaraman, C., Bernacki, K. & Jacobson, M. P.
17.
18.
19.
20.
21.
22.
23.
24.
25.
26.
27.
28.
29.
30.
31.
32.
33.
(2005). Virtual screening against highly charged active
sites: identifying substrates of alpha–beta barrel enzymes. Biochemistry, 44, 2059–2071.
Huang, N., Kalyanaraman, C., Irwin, J. J. & Jacobson,
M. P. (2006). Physics-based scoring of protein–ligand
complexes: enrichment of known inhibitors in largescale virtual screening. J. Chem. Inf. Model. 46, 243–253.
Perola, Emanuele (2006). Minimizing false positives in
kinase virtual screens. Proteins: Struct., Funct., Bioinf.
64, 422–435.
Lee, M. R. & Sun, Y. (2007). Improving docking accuracy through molecular mechanics generalized Born
optimization and scoring. J. Chem. Theory Comput. 3,
1106–1119.
Lyne, P. D., Lamb, M. L. & Saeh, J. C. (2006). Accurate
prediction of the relative potencies of members of a
series of kinase inhibitors using molecular docking
and MM–GBSA scoring. J. Med. Chem. 49, 4805–4808.
Mobley, D. L., Graves, A. P., Chodera, J. D.,
McReynolds, A. C., Shoichet, B. K. & Dill, K. A.
(2007). Predicting absolute ligand binding free energies to a simple model site. J. Mol. Biol. 371, 1118–1134.
Morton, A., Baase, W. A. & Matthews, B. W. (1995).
Energetic origins of specificity of ligand binding in an
interior nonpolar cavity of T4 lysozyme. Biochemistry,
34, 8564–8575.
Wei, B. Q., Baase, W. A., Weaver, L. H., Matthews,
B. W. & Shoichet, B. K. (2002). A model binding site
for testing scoring functions in molecular docking.
J. Mol. Biol. 322, 339–355.
Fitzgerald, M. M., Churchill, M. J., McRee, D. E. &
Goodin, D. B. (1994). Small molecule binding to an
artificially created cavity at the active site of cytochrome c peroxidase. Biochemistry, 33, 3807–3818.
Musah, R. A., Jensen, G. M., Bunte, S. W., Rosenfeld,
R. J. & Goodin, D. B. (2002). Artificial protein cavities
as specific ligand-binding templates: characterization
of an engineered heterocyclic cation-binding site that
preserves the evolved specificity of the parent protein.
J. Mol. Biol. 315, 845–857.
DeLano, W. L. (2002). The PyMOL Molecular Graphics
System, DeLano Scientific, San Carlos, CA.
Chang, C. E. & Gilson, M. K. (2004). Free energy,
entropy, and induced fit in host–guest recognition:
calculations with the second-generation mining minima algorithm. J. Am. Chem. Soc. 126, 13156–13164.
Brenk, R., Vetter, S. W., Boyce, S. E., Goodin, D. B. &
Shoichet, B. K. (2006). Probing molecular docking in
a charged model binding site. J. Mol. Biol. 357,
1449–1470.
Graves, A. P., Brenk, R. & Shoichet, B. K. (2005).
Decoys for docking. J. Med. Chem. 48, 3714–3728.
Wei, B. Q., Weaver, L. H., Ferrari, A. M., Matthews,
B. W. & Shoichet, B. K. (2004). Testing a flexiblereceptor docking algorithm in a model binding site.
J. Mol. Biol. 337, 1161–1182.
Deng, Y. & Roux, B. (2006). Calculation of standard
binding free energies: aromatic molecules in the T4
lysozyme L99A mutant. J. Chem. Theory Comput. 2,
1255–1273.
Boresch, S., Tettinger, F., Leitgeb, M. & Karplus, M.
(2003). Absolute binding free energies: a quantitative
approach for their calculation. J. Phys. Chem. B, 107,
9535–9551.
Machicado, C., Lopez-Llano, J., Cuesta-Lopez, S.,
Bueno, M. & Sancho, J. (2005). Design of ligand binding
to an engineered protein cavity using virtual screening
and thermal up-shift evaluation. J. Comp.-Aided Mol.
Des. 19, 421–443.
934
34. Rosenfeld, R., Goodsell, D., Musah, R., Garret, M.,
Goodin, D. & Olson, A. (2003). Automated docking
of ligands to an artificial active site: augmenting
crystallographic analysis with computer modeling.
J. Comput.-Aided Mol. Des, 17, 525–536.
35. Elowe, N. H., Blanchard, J. E., Cechetto, J. D. & Brown,
E. D. (2005). Experimental screening of dihydrofolate
reductase yields a “test set” of 50,000 small molecules
for a computational data-mining and docking competition. J. Biomol. Screening, 10, 653–657.
36. Jacobson, M. P., Pincus, D. L., Rapp, C. S., Day, T. J.,
Honig, B., Shaw, D. E. & Friesner, R. A. (2004). A
hierarchical approach to all-atom protein loop prediction. Proteins, 55, 351–367.
37. Jacobson, M. P., Kaminski, G. A., Friesner, R. A. &
Rapp, C. S. (2002). Force field validation using protein
side chain prediction. J. Phys. Chem. B, 106, 11673–11680.
38. Lorber, D. M. & Shoichet, B. K. (1998). Flexible ligand
docking using conformational ensembles. Protein Sci.
7, 938–950.
39. Irwin, J. J. & Shoichet, B. K. (2005). ZINC—a free
database of commercially available compounds for
virtual screening. J. Chem. Inf. Model. 45, 177–182.
40. Morton, A. & Matthews, B. W. (1995). Specificity of
ligand binding in a buried nonpolar cavity of T4
lysozyme: linkage of dynamics and structural plasticity. Biochemistry, 34, 8576–8588.
41. Mobley, D. L., Chodera, J. D. & Dill, K. A. (2007).
Confine-and-release method: obtaining correct binding free energies in the presence of protein conformational change. J. Chem. Theory Comput. 3, 1231–1235.
42. Song, L., Kalyanaraman, C., Fedorov, A. A., Fedorov,
E. V., Glasner, M. E., Brown, S. et al. (2007). Prediction
and assignment of function for a divergent N-succinyl
amino acid racemase. Nat. Chem. Biol. 3, 486–491.
43. Claussen, H., Buning, C., Rarey, M. & Lengauer, T.
(2001). FlexE: efficient molecular docking considering
protein structure variations. J. Mol. Biol. 308, 377–395.
44. Leach, A. R., Shoichet, B. K. & Peishoff, C. E. (2006).
Prediction of protein–ligand interactions. Docking and
scoring: successes and gaps. J. Med. Chem. 49, 5851–5855.
45. Li, J. B., Zhu, T. H., Cramer, C. J. & Truhlar, D. G. (1998).
New class IV charge model for extracting accurate
partial charges from wave functions. J. Phys. Chem. A,
102, 1820–1831.
46. Chambers, C. C., Hawkins, G. D., Cramer, C. J. &
Truhlar, D. G. (1996). Model for aqueous solvation
based on class IV atomic charges and first solvation
shell effects. J. Phys. Chem. 100, 16385–16398.
47. Banks, J. L., Beard, H. S., Cao, Y., Cho, A. E., Damm,
W., Farid, R. et al. (2005). Integrated Modeling Program, Applied Chemical Theory (IMPACT). J. Comput.
Chem. 26, 1752–1780.
48. Kaminski, G. A., Friesner, R. A., Tirado-Rives, J. &
Jorgensen, W. L. (2001). Evaluation and reparametrization of the OPLS-AA force field for proteins via
comparison with accurate quantum chemical calculations on peptides. J. Phys. Chem. B, 105, 6474–6487.
Rescoring Docking Hit Lists
49. Gallicchio, E., Zhang, L. Y. & Levy, R. M. (2002). The
SGB/NP hydration free energy model based on the
surface generalized born solvent reaction field and
novel nonpolar hydration free energy estimators.
J. Comput. Chem. 23, 517–529.
50. Zhu, K., Shirts, M. R., Friesner, R. A. & Jacobson, M. P.
(2007). Multiscale optimization of a truncated Newton
minimization algorithm and application to proteins
and protein–ligand complexes. J. Chem. Theory Comput. 3, 640–648.
51. Wang, J., Wang, W., Kollman, P. A. & Case, D. A.
(2006). Automatic atom type and bond type perception in molecular mechanical calculations. J. Mol.
Graphics Modell. 25, 247–260.
52. Wang, J., Wolf, R. M., Caldwell, J. W., Kollman, P. A. &
Case, D. A. (2004). Development and testing of a general amber force field. J. Comput. Chem. 25, 1157–1174.
53. Gasteiger, J. & Marsili, M. (1980). Iterative partial
equalization of orbital electronegativity: a rapid access
to atomic charges. Tetrahedron, 36, 3219–3288.
54. Jakalian, A., Jack, D. B. & Bayly, C. I. (2002). Fast,
efficient generation of high-quality atomic charges.
AM1-BCC model: II. Parameterization and validation.
J. Comput. Chem. 23, 1623–1641.
55. Bayly, C. I., Cieplak, P., Cornell, W. D. & Kollman,
P. A. (1993). A well-behaved electrostatic potential
based method using charge restraints for deriving
atomic charges—the RESP model. J. Phys. Chem. 97,
10269–10280.
56. Giammona, D. A. (1984). An Examination of Conformational Flexibility in Porphyrins and Bulky-Ligand Binding
in Myoglobin, Ph.D. Dissertation, University of California, Davis.
57. Onufriev, A., Bashford, D. & Case, D. A. (2004).
Exploring protein native states and large-scale conformational changes with a modified generalized born
model. Proteins: Struct., Funct., Bioinf. 55, 383–394.
58. Weiser, J., Shenkin, P. S. & Still, W. C. (1999). Approximate atomic surfaces from linear combinations
of pairwise overlaps (LCPO). J. Comput. Chem. 20,
217–230.
59. Kirchhoff, W. (1993). EXAM: a two-state thermodynamic
analysis program, NIST, Gaithersburg, MD.
60. Otwinowski, Z. & Minor, W. (1997). Processing of
X-ray diffraction data collected in oscillation mode. In
Methods in Enzymology (Carter, C. W. J. & Sweet, R. M.,
eds), Methods in Enzymology, vol. 276, pp. 307–326.
Academic Press, New York.
61. Schuttelkopf, A. W. & van Aalten, D. M. F. (2004).
PRODRG: a tool for high-throughput crystallography
of protein–ligand complexes. Acta Crystallogr., Sect. D:
Biol. Crystallogr. 60, 1355–1363.
62. Bailey, S. (1994). The CCP4 Suite—programs for
protein crystallography. Acta Crystallogr., Sect. D:
Biol. Crystallogr. 50, 760–763.
63. Emsley, P. & Cowtan, K. (2004). Coot: model-building
tools for molecular graphics. Acta Crystallogr., Sect. D:
Biol. Crystallogr. 60, 2126–2132.